Ex. this type of price estimation is quite naive - some models can require 2-3x the number of tokens to achieve the same level of intelligence. Artificial Analysis' own cost per task is a more fair estimation of cost.
https://artificialanalysis.ai/models#price-cost
https://artificialanalysis.ai/models?models=gpt-6-luna%2Cglm...
https://artificialanalysis.ai/models?cost=intelligence-vs-co...
Spark is such a bad model, you are far better off using Astra Medium than using Spark.
It's the equivalent of ai slop writing in a data format.
Depending on your memory, you'll need to use the weaker Q4 versions but they still perform well.
It ranks higher than GPT-5.3 Codex (xhigh) or Claude Opus 4.6 (max) so is great for pairing with https://github.com/kunchenguid/gnhf for nightly experimentation, cleanup, or recommendation lists for in the morning.
I've had more than a few people tell me "oh, it's so much cheaper to use a $20 claude account" or "i've never hit a limit ever using my openai". Inevitably.. I end up reading/hearing "oh, I need to give it another couple hours to start using it again"... I've never hit that with my approach, even if it's costing me a bit more. Being able to work when I want when I have time has some value.
I also have openai and anthropic direct API billing set up for hosted and client projects that need to call out to an LLM service.
If I want to spend $100 on LLMs next month, what should I do? Get Claude because Opus 5.5/Fable 5.1 are scoring well? Get Grok because 4.7 is supposedly a good mix of competence and cost? Try out a Chinese model? Don't do a subscription at all like this site is saying?
This is the one that works good for me on 32gb:
https://huggingface.co/ISTA-DASLab/Qwen3.8-27B-GSQ-RCO-GGUF
Specifically this one: Qwen3.8-27B-GSQ-RCO-IQ3_S-mtp.gguf
It would be handy to have a site like this one that takes into account the various deals and attempts to calculate the number of tokens per monthly fee for a chosen model. I realize this makes the task a lot more difficult.
2. It would be handy to have a chart like that for the AI hardware that people own. It helps you decide which model to run (resulting in different levels of intelligence and speed). Also difficult to please everyone (preprocessing vs token generation for example) and to keep updated!
I found https://llm-list.com/ yesterday and when I had a detailed look, I quickly found outdated entries, for example looking at GLM 5.3 flash it listed several providers as "free" that weren't free any longer.
For example I've been really enjoying Deepseek v4.1 Flash, it's very "straightforward" to the point of being almost dumb sometimes, but it's absolutely relentless and would solve almost any problem no matter how inefficient the solution is.
No idea how to measure all that, just average CoT length per task is probably a good approximation for some things, but not others.
It compares Artificial Analysis scores against usage limits on Opencode Go, so I can see where to waste my quota most efficiently. Updated whenever I feel like it.
Create your own, private evaluation system that reflects your use and your constraints. Focus on hard cases you encountered, find cases that break on one tier and not on another.
For example "Raw capability: Math Index", shows GPT-5.2 as highest. It does not even mention astra, of fable.
There are only 6 categories in the site, but a category entirely created out of thin air for Qwen3.8-Flash-Next?
Literally "$0.23 to under $0.24" means EXACTLY $0.23.
This is in the same ballpark for that same model on openrouter (https://openrouter.ai/qwen/qwen3.8-27b), highly dependent on input/output mix. Deepseek is a much more capable model that you can't run locally on normal hardware, and their rates are insanely cheap ($0.04 / $1 per 1M).
And so far I haven't considered the cost of the hardware. I happen to have a gaming PC that can be put to use on inference when not gaming, but given these numbers I don't think I would buy new hardware to do inference at home. Unless my math is wrong, it seems you're way better off paying for Deepseek than running Qwen or some other locally runnable model yourself. Of course if you have specific privacy requirements or prefer something unique about a particular model you can run locally, the equation changes.
[1] https://philippdubach.com/posts/jev-model-router-for-pi/
[dead]
[dead]
I often pair them up, and I have an Astra or Sol work as a supervisor and reviewer while Qwen 27B FP8 or Qwen 3.8 Flash Next implements things. I mostly do it as an experiment, just to see what kind of level of autonomy I can get, and they are slow to getting a decent reviewed outcome despite Qwen27B running at 100+ tps and 3.5-4K prefill rates and Qwen3.8 Next at 40 tps and 1-2K prefill. I've been also using similar approach more with OMP, not just the straight Pi harness. And OMP seems to be slower because it has more guardrails. OMP has an interesting feature where you can assign a better LLM as an advisor, wehere it just sort of monitors the progress and injects guidance. And it definitely helps, but one has to be careful. It actually turns out to be expensive if the cache reads are expensive. I learned it the hard way. Where on Fireworks' API, the cache rates for GLM 5.3 flash are quite a bit more expensive than for DeepSeek, and a simple runs ended up costing me three bucks in oversight. So a better way is to have a Frontier model running a separate tmux pane and just directing it to Wake up every 10 minutes, take a peek at what's going on, review the milestones give feedback and then sleep. This turns out to be pretty decent cost saving strategy when quote needs to be stretched. Paradoxically, OpenAI tightening up their quota allowance once they released Astra actually pushed me into all these sorts of experiments, and it's actually been interesting. I've been exploring all these smaller flash models, and it's been nice. I do like using local LLMs for chore type tasks that are just mostly information gathering, post-session reviews, stuff like that.
I find Qwen Flash Next quite competent as well. 27B is a solid worker like you said. If you batch work and let them crank they do remarkably well. I am working on.
I started using Herdr and taught my agents to use it. So I use OMP loop and or goal, and it has a review cycles to wake up an Opus or Sol reviewer to make sure nothing is going off the rails with a local qwen flash next coordinating for me. I kinda prefer Sol, it seems like a more patient and thorough model, especially Sol 6, but Opus 5.5 is really good and its voice and attitude is not as grating as Opus 5 for sure.
I have a few V100 GPU running Qwen 27B and they do all the work overnight. Not quite the same speeds you have yet, but this is V100 machine and an old gaming machine with 16GB 4080 and a handful of 32GB V100s... all in less than 3K (ignoring that my gaming machine is 3 years old, but runs qwen flash next for free now as I game not a lot) for my little "we have AI at home" projects and there is a lot of interest in these old GPU now because they are rolling out of data centers now.
For my local work and personal projects... they just seem to be getting done in this setup. Every few days I sit down and do a big cycle with astra/fable/opus batch things up. I have projects that are basically "i want to see what happens" to "I want this to be good, I understand the code". Some of the throwaway projects that have just kind of magically finished more or less how I wanted have been great.
GNHF seems exactly what I've been aiming for to handle overnight tasks.
Consider bumping reasoning down to Medium as a default though, I agree with simonw it over thinks https://simonwillison.net/2026/Aug/16/qwen-38-27b/
I guess I should still keep experimenting though. Maybe I'm just not using a dense model correctly.
Even on the original one, the authors have replied to such issues with working solutions: https://github.com/PrismML-Eng/Bonsai-demo/issues/131
The sheer number of geniuses working so fast on this repo never ceases to amaze me.
Would you mind sharing how much it's costing you? I've been wanting to use APIs from within Intellij, but I hesitate because of uncertainty about the cost.
Is it worth the ~10x extra cost over the subscriptions? (This is obviously a leading question). Also, I think you can use OpenAI's subcription login with Air, but not Claude's.
Junie is their agent(?) that can connect to multiple providers. AFAIK you can bring-your-own-key or buy a subscription from them that lets you use your credits across multiple vendors (openai, anthropic, google) across a handful of models from each vendor. So yeah, you can use your openai credentials in junie vs paying for jetbrains junie credits.
Is it 'worth it'? I dunno. I've not used dedicated subscriptions. I adopted Junie when it first came out, and it's been 'good enough'. As I say, I've never hit the 'oh, i have to stop for several hours until they let me in again' wall because of just using the API approach.
As my monthly spend goes up, I might try something else out, but... gpt6-luna just came out. That is likely going to cut my spend already.
I can't remember the last time any real recourse has mattered for companies getting breached or mishandling my data. Their stock just goes up and the govt just shrugs.
Then i got $20 into DeepSeek and i've been using those $20 for two weeks every day now. Use case is automating computer/browser use - Astra is really good at it, but very expensive, Sol and Luna haven't been that great at it, Deepseek as at about 80% of Astra but lasts forever.
Also for coding i use OMP directly with Deepseek.
It's been rather awesome because even maxed out $20 gets you very far in between Codex resets.
I also tried developing a small game where I have Astra use as many Deepseek workers and it needs and it only does the 3D objects and high level architecture and integration, but it goes way too quickly to see good results yet.
But I guess if I had an agent vibecoding on it's own, I'd go with subscription instantly.
Otherwise DeepSeek Flash 4.1 is dirt cheap (other "Flash" models are not that expensive either). I pay (very few dollars) out of my own pocket.
There are many things where having an API Key is necessary.
Maybe I’ve missed the boat though: is there now a method to use an api key to access a subscription?
It’s a good thing we have access to those very cheap and useful Chinese models.
which ones do you use?
- Qwen3.5 9B
- Qwen3.6 35B A3B
- Qwen3.8 27B
- Gemma 4 26B A4B
- Gemma 31B
- Muse Glimmer 30B
The MacMini is used solely for inference. llama.cpp + llama-swap.
People are saying local models are crap and serves no purposes.
I use them to help me spell check, write emails, write text messages, write JIRA tickets, write PR comments, etc.
I also use them in coding agents to complete different tasks.
I find them quite useful!
Do you think this will be a problem in the future?
The subscription from frontier labs can handle all that (though each has its best strenght, they all perform well on most things), but it's usually a bit of a waste compared to taking the best one for you.
And if you want Astra/Fable/Opus frontier level, then there's no option at all.
But if you don't need that, or you don't need speed... That opens up the discussion. I've been impressed even with how Siri's been doing with the Apple Foundation Models in MacOS/iOS 27 given how small they are.
Edit: I can't even fully spec the M5 Ultra Mac Studio you'd need for GLM5.3 Flash since 512GB isn't available yet, but it's already at $9500 for 256GB RAM.
It's probably worth letting the user specify their actual costs in such a tool. I run a Framework Desktop 128GB that I bought before memory prices got crazy; the current retail price is almost double what I actually paid a year ago.
You will be able to run some pretty impressive models locally. Good for you.